Updated November 2026 — Steal math and fio percentiles below were collected on a 4 vCPU KVM guest on an AMD EPYC host, kernel 6.8, during a 90-minute fio randread/randwrite mixed run. The point is to turn marketing tiers into numbers you can verify in a terminal.
A VPS tier is sold as a set of numbers: vCPU, RAM, disk GB, “NVMe”, “unmetered”. None of those describe what you actually receive. What you receive is a share of a host CPU, a share of a block device queue, and whatever the hypervisor decides to hand you when the host is busy. Two counters tell you the truth: steal time from /proc/stat, and latency percentiles from fio. Everything else is inference.
This is a verification procedure, not a benchmark-for-fun exercise. Run it in the first 48 hours of a new tier, and again after any host migration, so you have a before and after. If you have not picked the tier yet, the cloud VPS benefits page covers what the isolation model does and does not guarantee.
Steal Time: The Exact Math
The CPU line in /proc/stat has ten fields in kernel 6.8: user, nice, system, idle, iowait, irq, softirq, steal, guest, guest_nice. Steal is field 9. Total is the sum of fields 1 through 8 (user..steal); guest time is already included in user and system, so adding fields 9 and 10 double counts.
# steal over a 10 s window, percentage of guest-visible CPU time
read -r a < /proc/stat; sleep 10; read -r b < /proc/stat
python3 -c "a='$a'.split(); b='$b'.split(); t=sum(int(b[i])-int(a[i]) for i in range(1,9)); print('steal %.2f%%' % (100*(int(b[8])-int(a[8]))/t))"
On a 4 vCPU guest the counter aggregates all four CPUs only if you read /proc/stat on the first line; the per-CPU lines follow. A single hot vCPU being stolen from shows up in its own line. Always read the aggregate for a headline number and the per-CPU lines when the aggregate looks fine but one thread is slow.
Ten seconds is too short to be meaningful. Sample for at least ten minutes, and compute the median of 60 samples, not the mean, because one 200 ms host migration skews a mean badly. A tier that advertises "dedicated cores" should sit under 0.5% steal; a shared tier with a fair-share scheduler often sits at 1-3% at idle.
Disk Latency Percentiles That Actually Bite
Average latency hides everything that matters. Run fio with percentile reporting and a realistic random mix, and compare p50, p99 and p99.9. For a database the p99.9 is what shows up in your slow query log at 3 a.m.
fio --name=rw --filename=/dev/nvme0n1 --direct=1 \
--rw=randrw --rwmixread=70 --bs=4k --ioengine=libaio \
--iodepth=16 --numjobs=1 --runtime=300 --time_based \
--percentile_list=50:90:99:99.9 --group_reporting
The percentile_list prints one row per percentile in the clat section. On a healthy NVMe tier expect p50 around 80-120 µs, p99 around 600 µs to 1 ms, and p99.9 under 4 ms at depth 16. Numbers above those are not "wrong", but they need an explanation — and the explanation is either the device, the queue, or steal.
| Metric | Healthy NVMe tier | Shared tier, healthy | Red flag |
|---|---|---|---|
| p50 read | 80-120 µs | 150-250 µs | > 400 µs |
| p99 read | 0.6-1 ms | 2-4 ms | > 10 ms |
| p99.9 read | < 4 ms | < 15 ms | > 40 ms |
| steal (10 min median) | < 0.5% | 1-3% | > 5% |
| iostat r_await | < 1.5 ms | < 4 ms | > 10 ms |
Cross-Checking Steal Against Latency
Steal and latency must be read together, because the same latency curve can mean two things. Collect both at once: run the steal sampler in one pane and fio in another, with identical start and end times. Then bucket the fio samples by the steal level of the minute they landed in, and you get a latency-per-steal-level table instead of two unrelated numbers.
The comparison is what turns a feeling into a claim. If latency rises 15% for every extra point of steal, the host is the cause. If latency rises while steal is flat, the cause is the device or the queue. Keep both series in one CSV with a timestamp column so the join is trivial when the ticket is open.
- Steal low, latency high → device or queue; a host migration may not help.
- Steal high, latency high, both rising together → CPU contention from a noisy vCPU neighbour.
- Steal high, latency normal → workload is I/O bound; you are not actually harmed yet.
- Steal low, latency low, throughput low → queue depth limit; raise iodepth or move tiers.
The Decision Table
The whole point of the exercise is a defensible answer to one question: is this tier what was sold? The table below is the one to keep on the wall. It maps the measured combination to a verdict and a next action, so a support ticket can cite numbers instead of a feeling.
The verdict column has to be about structure, not speed. A tier that gives steal under 1% but caps the disk queue at depth 4 is behaving exactly as sold: the queue limit was in the fine print. The question is whether that structure supports your workload, and whether the provider described it honestly. A tier with low steal but 40 ms p99.9 is not behaving as sold, because nothing in the listing promised a device that stalls.
The same logic applies to CPU. An advertised 4 vCPU tier under a fair-share scheduler shows 1-3% steal at idle, and that is normal, not a fault. Sustained steal above 5% for more than fifteen minutes is where a ticket becomes worth filing: at that level you are paying for cores you do not receive, and any mildly CPU-sensitive workload can see it.
| Observed | Verdict | Action |
|---|---|---|
| Steal < 1%, p99 < 2x published | Tier is as sold | Keep; re-check quarterly |
| Steal 1-5%, p99 2-5x | Shared tier behaving normally | Tune depth; consider a higher tier |
| Steal > 5% sustained > 15 min | CPU over-subscribed | Ticket with /proc/stat samples |
| Listed IOPS unreachable at any depth | Queue limit or throttling | Ticket with fio JSON + iostat |
| p99.9 > 40 ms with low steal | Device failing or moving data | Escalate; do not tune around it |
How to Present the Evidence
A support ticket that says "the disk is slow" gets closed. A ticket with a fio JSON, a 10-minute steal series saved as CSV, and the exact minute where p99.9 crossed 40 ms gets escalated. Keep all three artefacts in the same directory with the run date, and keep the raw fio output rather than only the summary — engineers can rerun the analysis, they cannot rerun your deleted file.
If the verdict is "the tier is not as sold" and the response is "upgrade", the upgrade is only worth it if the new tier changes the structural limits: more queues, a dedicated device, or a pinning policy. That difference is exactly what managed vs unmanaged hosting discussions tend to skip past — the isolation model matters more than the brand.
Keep the verification cheap and repeatable. Ten minutes of fio and a one-line awk steal sampler, run monthly and stored, gives you a history you can compare against. Most complaints that "the VPS got slower" are true, and the proof is usually already on disk — you just have to be collecting it before the incident.




Leave a Reply
You must be logged in to post a comment.